Hoshipu/behavior-1k-embodiedAI-demos

BEHAVIOR-1K EmbodiedAI Demos - Task 0

This dataset contains categorized demonstrations for Task 0 (turning on radio) from the BEHAVIOR-1K dataset.

Dataset Structure

The dataset is organized into two main categories:

success-demos/

• 459 episodes of successful task completions
• 912,371 frames total
• Includes 200 episodes from original BEHAVIOR-1K demos (episodes 0-199)
• Includes 259 episodes from collected runs (episodes 5000-5258)

failure-demos/

• 533 episodes of failed task attempts
• 2,291,900 frames total
• Collected from various failure modes during data collection

Data Format

Each category follows the standard BEHAVIOR-1K structure:

<category>/
├── data/
│   └── task-0000/
│       └── episode_*.parquet
├── videos/
│   └── task-0000/
│       ├── observation.images.rgb.head/
│       ├── observation.images.rgb.left_wrist/
│       └── observation.images.rgb.right_wrist/
├── meta/
│   ├── info.json
│   ├── episodes.jsonl
│   └── tasks.jsonl
└── annotations/
    └── task-0000/

Meta Files

• info.json: Overall dataset metadata (total episodes, frames, features, etc.)
• episodes.jsonl: One JSON line per episode with episode_index, tasks, length, and metrics
• tasks.jsonl: Task definitions (task_index, task_name, task description)

Usage

from datasets import load_dataset

# Load success demos
success_ds = load_dataset("Hoshipu/behavior-1k-embodiedAI-demos", data_dir="success-demos")

# Load failure demos
failure_ds = load_dataset("Hoshipu/behavior-1k-embodiedAI-demos", data_dir="failure-demos")

Dataset Statistics

• Total Episodes: 992 (459 success + 533 failure)
• Total Frames: 3,204,271
• FPS: 30
• Robot Type: R1Pro
• Codebase Version: v2.1
• Task: Turn on the radio receiver that's on the table in the living room

Video Streams

Each episode includes three RGB camera views:

• observation.images.rgb.head: 720x720 head-mounted camera
• observation.images.rgb.left_wrist: 480x480 left wrist camera
• observation.images.rgb.right_wrist: 480x480 right wrist camera

Citation

If you use this dataset, please cite the original BEHAVIOR-1K paper:

@article{behavior1k2025,
  title={BEHAVIOR-1K: A Human-Centered, Embodied AI Benchmark with 1,000 Everyday Activities and Realistic Simulation},
  author={...},
  journal={...},
  year={2025}
}